The ecosystem of AI agent tools for development has grown so large that it's easy to burn a day choosing between similar projects. Below are six repositories covering different layers of the job: from writing code autonomously to orchestrating several agents and ready-made integrations via MCP.
How this roundup is organized
The six projects fall into three groups by their role in the workflow: code agents (OpenHands, Aider, Task Master) work directly with your repository; orchestration (CrewAI, LangGraph, Dify) is what you need when one model can't finish the job and you need several roles or state between sessions; MCP servers (awesome-mcp-servers) are ready integrations between an agent and external systems instead of a home-made connector.
OpenHands — an autonomous developer agent
OpenHands writes code, runs terminal commands, fixes bugs and works with the browser without a human watching over every step.
It comes in three formats: as a CLI tool, a local web interface or a cloud service, with support for different LLMs — it isn't tied to a single model.
Repository: github.com/OpenHands/openhands
Aider — pair programming in the terminal
Aider is an AI partner that works right in the terminal, alongside a developer's normal workflow, with no separate IDE or web interface.
It builds a map of the whole codebase before proposing changes, understands over 100 programming languages and commits edits to Git itself with meaningful messages — so the history is actually readable afterwards instead of a guessing game.
Repository: github.com/aider-ai/aider
Task Master — task management for AI-driven development
Task Master plugs into Cursor, Windsurf and other editors and solves a specific problem: backlog chaos when tasks for the agent are made up on the fly and nobody tracks the dependencies between them.
It breaks requirements down into tasks, tracks the links between them and suggests what to do next — useful when more than one person is working on a project with an agent.
Repository: github.com/eyaltoledano/claude-task-master
CrewAI — a team of agents with different roles
CrewAI runs several AI agents with clearly separated roles that solve one task together — one agent researches, a second writes, a third reviews, for example.
It doesn't depend on LangChain and works with any LLM, including local models via Ollama — so you aren't locked into a particular cloud provider.
Repository: github.com/crewaiinc/crewai
LangGraph and Dify — agents that outlive a single session
LangGraph is a low-level framework for agents with complex logic: it can survive failures and resume where it stopped, keeping memory between sessions. It's used in production by large companies, not only in pet projects.
Dify solves a similar problem in a different way — it wraps agent building in a visual workflow builder with RAG support and Docker deployment, so you can assemble similar logic without writing the orchestration code by hand.
Repositories: github.com/langchain-ai/langgraph, github.com/langgenius/dify
awesome-mcp-servers — ready integrations instead of home-made ones
A catalog of ready MCP servers for connecting an agent to external systems: databases, files, APIs, cloud services — split into categories.
Before writing your own connector for an agent, it's worth checking the catalog: the integration you need has often already been implemented and is maintained by the community.
Repository: github.com/punkpeye/awesome-mcp-servers
How to choose between them
If the job is to work with existing code at your own pace, start with Aider (terminal, full control) or OpenHands (more autonomy). If the project needs several roles or has to survive a failure without losing context — look at CrewAI, LangGraph or Dify depending on whether you want a visual builder. Task Master and awesome-mcp-servers aren't alternatives to an agent but infrastructure around it: the first organizes tasks, the second provides ready integrations.
How we do it
We test tools like these on our own tasks — the specific prompts for integrating each agent and the setup gotchas are collected in the knowledge base rather than retold from the documentation. The qvib engine for generating code and websites is built on similar principles of agent orchestration and MCP integrations — the same pipeline as in the knowledge base.
Related links
- GitHub tool roundups: qvib.pro/arsenal/github-podborki/
- Developer tools: qvib.pro/arsenal/instrumenty-dlya-razrabotchika/
- The "Claude skills" section: qvib.pro/arsenal/claude-skills/